bool RuleTableLoaderStandard::Load(FormatType format
, const std::vector<FactorType> &input
, const std::vector<FactorType> &output
, const std::string &inFile
, const std::vector<float> &weight
, size_t /* tableLimit */
, const LMList &languageModels
, const WordPenaltyProducer* wpProducer
, RuleTableTrie &ruleTable)
{
PrintUserTime(string("Start loading text SCFG phrase table. ") + (format==MosesFormat?"Moses ":"Hiero ") + " format");
const StaticData &staticData = StaticData::Instance();
const std::string& factorDelimiter = staticData.GetFactorDelimiter();
string lineOrig;
size_t count = 0;
std::ostream *progress = NULL;
IFVERBOSE(1) progress = &std::cerr;
util::FilePiece in(inFile.c_str(), progress);
// reused variables
vector<float> scoreVector;
StringPiece line;
std::string hiero_before, hiero_after;
while(true) {
try {
line = in.ReadLine();
} catch (const util::EndOfFileException &e) { break; }
if (format == HieroFormat) { // inefficiently reformat line
hiero_before.assign(line.data(), line.size());
ReformatHieroRule(hiero_before, hiero_after);
line = hiero_after;
}
util::TokenIter<util::MultiCharacter> pipes(line, "|||");
StringPiece sourcePhraseString(*pipes);
StringPiece targetPhraseString(*++pipes);
StringPiece scoreString(*++pipes);
StringPiece alignString(*++pipes);
// TODO(bhaddow) efficiently handle default instead of parsing this string every time.
StringPiece ruleCountString = ++pipes ? *pipes : StringPiece("1 1");
if (++pipes) {
stringstream strme;
strme << "Syntax error at " << ruleTable.GetFilePath() << ":" << count;
UserMessage::Add(strme.str());
abort();
}
bool isLHSEmpty = (sourcePhraseString.find_first_not_of(" \t", 0) == string::npos);
if (isLHSEmpty && !staticData.IsWordDeletionEnabled()) {
TRACE_ERR( ruleTable.GetFilePath() << ":" << count << ": pt entry contains empty target, skipping\n");
continue;
}
scoreVector.clear();
for (util::TokenIter<util::AnyCharacter, true> s(scoreString, " \t"); s; ++s) {
char *err_ind;
scoreVector.push_back(strtod(s->data(), &err_ind));
UTIL_THROW_IF(err_ind == s->data(), util::Exception, "Bad score " << *s << " on line " << count);
}
const size_t numScoreComponents = ruleTable.GetFeature()->GetNumScoreComponents();
if (scoreVector.size() != numScoreComponents) {
stringstream strme;
strme << "Size of scoreVector != number (" << scoreVector.size() << "!="
<< numScoreComponents << ") of score components on line " << count;
UserMessage::Add(strme.str());
abort();
}
// parse source & find pt node
// constituent labels
Word sourceLHS, targetLHS;
// source
Phrase sourcePhrase( 0);
sourcePhrase.CreateFromStringNewFormat(Input, input, sourcePhraseString, factorDelimiter, sourceLHS);
// create target phrase obj
TargetPhrase *targetPhrase = new TargetPhrase(Output);
targetPhrase->CreateFromStringNewFormat(Output, output, targetPhraseString, factorDelimiter, targetLHS);
targetPhrase->SetSourcePhrase(sourcePhrase);
// rest of target phrase
targetPhrase->SetAlignmentInfo(alignString, sourcePhrase);
targetPhrase->SetTargetLHS(targetLHS);
targetPhrase->SetRuleCount(ruleCountString, scoreVector[0]);
//targetPhrase->SetDebugOutput(string("New Format pt ") + line);
// component score, for n-best output
std::transform(scoreVector.begin(),scoreVector.end(),scoreVector.begin(),TransformScore);
std::transform(scoreVector.begin(),scoreVector.end(),scoreVector.begin(),FloorScore);
targetPhrase->SetScoreChart(ruleTable.GetFeature(), scoreVector, weight, languageModels,wpProducer);
TargetPhraseCollection &phraseColl = GetOrCreateTargetPhraseCollection(ruleTable, sourcePhrase, *targetPhrase, sourceLHS);
phraseColl.Add(targetPhrase);
count++;
}
// sort and prune each target phrase collection
SortAndPrune(ruleTable);
return true;
}
bool RuleTrieLoader::Load(const std::vector<FactorType> &input,
const std::vector<FactorType> &output,
const std::string &inFile,
const RuleTableFF &ff,
RuleTrie &trie)
{
PrintUserTime(std::string("Start loading text phrase table. Moses format"));
const StaticData &staticData = StaticData::Instance();
// const std::string &factorDelimiter = staticData.GetFactorDelimiter();
std::size_t count = 0;
std::ostream *progress = NULL;
IFVERBOSE(1) progress = &std::cerr;
util::FilePiece in(inFile.c_str(), progress);
// reused variables
std::vector<float> scoreVector;
StringPiece line;
double_conversion::StringToDoubleConverter converter(double_conversion::StringToDoubleConverter::NO_FLAGS, NAN, NAN, "inf", "nan");
while(true) {
try {
line = in.ReadLine();
} catch (const util::EndOfFileException &e) {
break;
}
util::TokenIter<util::MultiCharacter> pipes(line, "|||");
StringPiece sourcePhraseString(*pipes);
StringPiece targetPhraseString(*++pipes);
StringPiece scoreString(*++pipes);
StringPiece alignString;
if (++pipes) {
StringPiece temp(*pipes);
alignString = temp;
}
if (++pipes) {
StringPiece str(*pipes); //counts
}
bool isLHSEmpty = (sourcePhraseString.find_first_not_of(" \t", 0) == std::string::npos);
if (isLHSEmpty && !staticData.IsWordDeletionEnabled()) {
TRACE_ERR( ff.GetFilePath() << ":" << count << ": pt entry contains empty target, skipping\n");
continue;
}
scoreVector.clear();
for (util::TokenIter<util::AnyCharacter, true> s(scoreString, " \t"); s; ++s) {
int processed;
float score = converter.StringToFloat(s->data(), s->length(), &processed);
UTIL_THROW_IF2(std::isnan(score), "Bad score " << *s << " on line " << count);
scoreVector.push_back(FloorScore(TransformScore(score)));
}
const std::size_t numScoreComponents = ff.GetNumScoreComponents();
if (scoreVector.size() != numScoreComponents) {
UTIL_THROW2("Size of scoreVector != number (" << scoreVector.size() << "!="
<< numScoreComponents << ") of score components on line " << count);
}
// parse source & find pt node
// constituent labels
Word *sourceLHS = NULL;
Word *targetLHS;
// create target phrase obj
TargetPhrase *targetPhrase = new TargetPhrase(&ff);
// targetPhrase->CreateFromString(Output, output, targetPhraseString, factorDelimiter, &targetLHS);
targetPhrase->CreateFromString(Output, output, targetPhraseString, &targetLHS);
// source
Phrase sourcePhrase;
// sourcePhrase.CreateFromString(Input, input, sourcePhraseString, factorDelimiter, &sourceLHS);
sourcePhrase.CreateFromString(Input, input, sourcePhraseString, &sourceLHS);
// rest of target phrase
targetPhrase->SetAlignmentInfo(alignString);
targetPhrase->SetTargetLHS(targetLHS);
//targetPhrase->SetDebugOutput(string("New Format pt ") + line);
if (++pipes) {
StringPiece sparseString(*pipes);
targetPhrase->SetSparseScore(&ff, sparseString);
}
if (++pipes) {
StringPiece propertiesString(*pipes);
targetPhrase->SetProperties(propertiesString);
}
targetPhrase->GetScoreBreakdown().Assign(&ff, scoreVector);
targetPhrase->EvaluateInIsolation(sourcePhrase, ff.GetFeaturesToApply());
TargetPhraseCollection &phraseColl = GetOrCreateTargetPhraseCollection(
trie, *sourceLHS, sourcePhrase);
phraseColl.Add(targetPhrase);
// not implemented correctly in memory pt. just delete it for now
delete sourceLHS;
count++;
}
// sort and prune each target phrase collection
if (ff.GetTableLimit()) {
SortAndPrune(trie, ff.GetTableLimit());
}
return true;
}
bool RuleTableLoaderStandard::Load(FormatType format
, const std::vector<FactorType> &input
, const std::vector<FactorType> &output
, const std::string &inFile
, size_t /* tableLimit */
, RuleTableTrie &ruleTable)
{
PrintUserTime(string("Start loading text SCFG phrase table. ") + (format==MosesFormat?"Moses ":"Hiero ") + " format");
const StaticData &staticData = StaticData::Instance();
const std::string& factorDelimiter = staticData.GetFactorDelimiter();
string lineOrig;
size_t count = 0;
std::ostream *progress = NULL;
IFVERBOSE(1) progress = &std::cerr;
util::FilePiece in(inFile.c_str(), progress);
// reused variables
vector<float> scoreVector;
StringPiece line;
std::string hiero_before, hiero_after;
double_conversion::StringToDoubleConverter converter(double_conversion::StringToDoubleConverter::NO_FLAGS, NAN, NAN, "inf", "nan");
while(true) {
try {
line = in.ReadLine();
} catch (const util::EndOfFileException &e) {
break;
}
if (format == HieroFormat) { // inefficiently reformat line
hiero_before.assign(line.data(), line.size());
ReformatHieroRule(hiero_before, hiero_after);
line = hiero_after;
}
util::TokenIter<util::MultiCharacter> pipes(line, "|||");
StringPiece sourcePhraseString(*pipes);
StringPiece targetPhraseString(*++pipes);
StringPiece scoreString(*++pipes);
StringPiece alignString;
if (++pipes) {
StringPiece temp(*pipes);
alignString = temp;
}
if (++pipes) {
StringPiece str(*pipes); //counts
}
bool isLHSEmpty = (sourcePhraseString.find_first_not_of(" \t", 0) == string::npos);
if (isLHSEmpty && !staticData.IsWordDeletionEnabled()) {
TRACE_ERR( ruleTable.GetFilePath() << ":" << count << ": pt entry contains empty target, skipping\n");
continue;
}
scoreVector.clear();
for (util::TokenIter<util::AnyCharacter, true> s(scoreString, " \t"); s; ++s) {
int processed;
float score = converter.StringToFloat(s->data(), s->length(), &processed);
UTIL_THROW_IF(isnan(score), util::Exception, "Bad score " << *s << " on line " << count);
scoreVector.push_back(FloorScore(TransformScore(score)));
}
const size_t numScoreComponents = ruleTable.GetNumScoreComponents();
if (scoreVector.size() != numScoreComponents) {
stringstream strme;
strme << "Size of scoreVector != number (" << scoreVector.size() << "!="
<< numScoreComponents << ") of score components on line " << count;
UserMessage::Add(strme.str());
abort();
}
// parse source & find pt node
// constituent labels
Word *sourceLHS;
Word *targetLHS;
// create target phrase obj
TargetPhrase *targetPhrase = new TargetPhrase();
targetPhrase->CreateFromString(Output, output, targetPhraseString, factorDelimiter, &targetLHS);
// source
Phrase sourcePhrase;
sourcePhrase.CreateFromString(Input, input, sourcePhraseString, factorDelimiter, &sourceLHS);
// rest of target phrase
targetPhrase->SetAlignmentInfo(alignString);
targetPhrase->SetTargetLHS(targetLHS);
//targetPhrase->SetDebugOutput(string("New Format pt ") + line);
if (++pipes) {
StringPiece sparseString(*pipes);
targetPhrase->SetSparseScore(&ruleTable, sparseString);
}
if (++pipes) {
StringPiece propertiesString(*pipes);
targetPhrase->SetProperties(propertiesString);
}
targetPhrase->GetScoreBreakdown().Assign(&ruleTable, scoreVector);
targetPhrase->Evaluate(sourcePhrase, ruleTable.GetFeaturesToApply());
TargetPhraseCollection &phraseColl = GetOrCreateTargetPhraseCollection(ruleTable, sourcePhrase, *targetPhrase, sourceLHS);
phraseColl.Add(targetPhrase);
count++;
}
// sort and prune each target phrase collection
SortAndPrune(ruleTable);
return true;
}
bool PhraseDictionaryMemory::Load(const std::vector<FactorType> &input
, const std::vector<FactorType> &output
, const string &filePath
, const vector<float> &weight
, size_t tableLimit
, const LMList &languageModels
, float weightWP)
{
const_cast<LMList&>(languageModels).InitializeBeforeSentenceProcessing();
const StaticData &staticData = StaticData::Instance();
m_tableLimit = tableLimit;
util::FilePiece inFile(filePath.c_str(), staticData.GetVerboseLevel() >= 1 ? &std::cerr : NULL);
size_t line_num = 0;
size_t numElement = NOT_FOUND; // 3=old format, 5=async format which include word alignment info
const std::string& factorDelimiter = staticData.GetFactorDelimiter();
Phrase sourcePhrase(0);
std::vector<float> scv;
scv.reserve(m_numScoreComponent);
TargetPhraseCollection *preSourceNode = NULL;
std::string preSourceString;
while(true) {
++line_num;
StringPiece line;
try {
line = inFile.ReadLine();
} catch (util::EndOfFileException &e) {
break;
}
util::TokenIter<util::MultiCharacter> pipes(line, util::MultiCharacter("|||"));
StringPiece sourcePhraseString(GrabOrDie(pipes, filePath, line_num));
StringPiece targetPhraseString(GrabOrDie(pipes, filePath, line_num));
StringPiece scoreString(GrabOrDie(pipes, filePath, line_num));
bool isLHSEmpty = !util::TokenIter<util::AnyCharacter, true>(sourcePhraseString, util::AnyCharacter(" \t"));
if (isLHSEmpty && !staticData.IsWordDeletionEnabled()) {
TRACE_ERR( filePath << ":" << line_num << ": pt entry contains empty source, skipping\n");
continue;
}
//target
std::auto_ptr<TargetPhrase> targetPhrase(new TargetPhrase());
targetPhrase->CreateFromString(output, targetPhraseString, factorDelimiter);
scv.clear();
for (util::TokenIter<util::AnyCharacter, true> token(scoreString, util::AnyCharacter(" \t")); token; ++token) {
char *err_ind;
// Token is always delimited by some form of space. Also, apparently strtod is portable but strtof isn't.
scv.push_back(FloorScore(TransformScore(static_cast<float>(strtod(token->data(), &err_ind)))));
if (err_ind == token->data()) {
stringstream strme;
strme << "Bad number " << token << " on line " << line_num;
UserMessage::Add(strme.str());
abort();
}
}
if (scv.size() != m_numScoreComponent) {
stringstream strme;
strme << "Size of scoreVector != number (" <<scv.size() << "!=" <<m_numScoreComponent<<") of score components on line " << line_num;
UserMessage::Add(strme.str());
abort();
}
size_t consumed = 3;
if (pipes) {
targetPhrase->SetAlignmentInfo(*pipes++);
++consumed;
}
ScoreComponentCollection sparse;
if (pipes) pipes++; //counts
if (pipes) {
//sparse features
SparsePhraseDictionaryFeature* spdf =
GetFeature()->GetSparsePhraseDictionaryFeature();
if (spdf) {
sparse.Assign(spdf,(pipes++)->as_string());
}
}
// scv good to go sir!
targetPhrase->SetScore(m_feature, scv, sparse, weight, weightWP, languageModels);
// Check number of entries delimited by ||| agrees across all lines.
for (; pipes; ++pipes, ++consumed) {}
if (numElement != consumed) {
if (numElement == NOT_FOUND) {
numElement = consumed;
} else {
stringstream strme;
strme << "Syntax error at " << filePath << ":" << line_num;
UserMessage::Add(strme.str());
abort();
}
}
//TODO: Would be better to reuse source phrases, but ownership has to be
//consistent across phrase table implementations
sourcePhrase.Clear();
sourcePhrase.CreateFromString(input, sourcePhraseString, factorDelimiter);
//Now that the source phrase is ready, we give the target phrase a copy
targetPhrase->SetSourcePhrase(sourcePhrase);
if (preSourceString == sourcePhraseString && preSourceNode) {
preSourceNode->Add(targetPhrase.release());
} else {
preSourceNode = CreateTargetPhraseCollection(sourcePhrase);
preSourceNode->Add(targetPhrase.release());
preSourceString.assign(sourcePhraseString.data(), sourcePhraseString.size());
}
}
// sort each target phrase collection
m_collection.Sort(m_tableLimit);
/* // TODO ASK OLIVER WHY THIS IS NEEDED
const_cast<LMList&>(languageModels).CleanUpAfterSentenceProcessing();
*/
return true;
}